Phil Bull ([email protected])
November 2020
Cosmology Fisher forecasting code for HI (21cm) intensity mapping experiments and spectroscopic galaxy surveys.
Released under the Academic Free License (AFL-3.0).
RadioFisher is a Fisher forecasting code for cosmology with intensity maps of the redshifted 21cm emission line of neutral hydrogen. The formalism implemented by this code is described in Bull, Ferreira, Patel and Santos (2014). It's written in Python, and makes heavy use of NumPy, SciPy, and matplotlib. You also need CAMB. I'm running it very happily on Ubuntu 20.04 and other Linux machines. It should also run fine on Macs.
The code is provided openly for inspection and re-use. If you use it, please cite us! If you have problems getting it to work, or have bugfixes, comments, or suggestions, please get in touch with me. This is an actively-used scientific code, so expect to have to get your hands dirty!
- Python (tested with 3.7)
- Recent NumPy and SciPy
- matplotlib
- mpi4py
- CAMB (http://camb.info/)
This is a collection of Python scripts, none of which need to be compiled or installed before use. However, you should compile CAMB and change the CAMB_EXEC variable in baofisher.py to point at this executable.
To get started, check out the git repository and make sure you have Python 2.7 and up-to-date versions of SciPy, NumPy, and matplotlib installed. You should also download and compile CAMB. It also helps if you install mpi4py too; although most of the Fisher forecasting code in baofisher.py doesn't need MPI, the full_experiment.py frontend code does. Finally, in order to use some of the existing experiment definitions, you’ll need to download the interferometer baselines package* and unpack it in the array_config/ subdirectory. *[http://philbull.com/radiofisher_array_config.tar.gz; 5 MB]
Next, edit the CAMB_EXEC variable at the top of baofisher.py to point at the directory where the CAMB executable resides (not the CAMB executable itself). At the moment, you also need to create a subdirectory called output/.
Now, to run your first forecast, call mpirun -n 1 ./full_experiment.py 0. This
will run a forecast for the first experiment listed in full_experiment.py using
only one processor, and it will take quite a while (several minutes, usually).
This is because the first time you run it, CAMB runs and produces a
high-resolution P(k) for your fiducial cosmology. This is cached, so subsequent
runs are much faster.
Output is stored in the output/ subdirectory. Look inside full_experiment.py to see what's saved; it's mostly a Fisher matrix per redshift bin (variable names are in the header of each file), and some auxiliary information about redshift bins, functions of redshift, and the binning of P(k) in k space.
The next time you run, try changing the first and second numbers, i.e.
mpirun -n <nproc> ./full_experiment.py <experiment-id>. Each redshift bin is
processed as a separate unit. Redshift bins are divided up between the
available CPUs. There is no point setting higher than the number of
redshift bins for a given experiment.
To process the output of the forecasts, try one of the plot_*.py scripts. You'll probably have to edit it first to specify which experiments you want to see (lists of experiments are always specified near the top of the script).
Here are some key files that you should know about. Make sure you look inside them; they're heavily commented and pretty much everything has a Python docstring.
- baofisher.py: Most of the forecasting code and a large number of helper functions are kept here.
- full_experiment.py: Script for running a full forecast for a given experiment.
- experiments.py: Specifications for a large number of experiments are defined here, as well as survey parameters and the fiducial cosmological parameters. Some of them need auxiliary baseline distribution files; these will be made public soon, but in the meantime just email me if you need them.
- galaxy_full_experiment.py: Runs a forecast for a galaxy redshift survey rather than an IM survey.
- plot_dlogp.py: Plots constraints on P(k) for given experiments. This is a useful first plotting script to use.
- plot_w0wa.py: Another useful plotting script to look at, as it works with cosmological parameters rather than just P(k) on its own.
As mentioned above, there is no point using more processors than an experiment has redshift bins, since the extra processors will just sit around idle.
You may find that your installation of SciPy/NumPy or CAMB causes some weird
behaviour and hogs a lot of processor sometimes when this code is running. This
might be caused by a conflict between MPI (which Radio Fisher uses) and OpenMP
(which some of the libraries just mentioned have optional support for). Try
calling the script with OMP_NUM_THREADS=2 or some other low number, e.g.
OMP_NUM_THREADS=2 mpirun -n 20 ./full_experiment.py 4.
If you use this code in a scientific work, please cite us! The relevant paper is the following:
Philip Bull, Pedro G. Ferreira, Prina Patel, and Mario Santos, ApJ 803, 21 (2015) [arXiv:1405.1452] [doi:10.1088/0004-637X/803/1/21].
Email me ([email protected]) with bug reports, patches, requests for features and so on. I'll be happy to help/fix things.